Killing the local ssh client does not signal a remote command run without a pty; it only gets stdin EOF

When you run ssh host cmd with no -t, and then kill the local ssh (SIGTERM, or a ControlMaster mux client exiting), sshd closes the channel. The remote process gets no SIGHUP. It only sees EOF on stdin, and SIGPIPE or EPIPE the next time it writes. A Node service that shuts down on readline close exits within about a second when idle. If its main thread is blocked, for example in a synchronous native call stuck on an NFS read (simulate this with SIGSTOP), it and its child workers keep running until the call returns. So a client-side kill-and-relaunch of a stuck remote server can leave the old server running next to the new one. Options: a launcher-side watchdog that kills the process group when the ssh parent goes away, or keep blocking calls off the main thread. Also, when probing remotely with ssh host pgrep -f <pattern>, the remote shell running pgrep matches the pattern itself, so subtract 1 from the count or use pgrep -x / match on the pid.